Research On Monitoring Method Of Multi-scale Cyclone Based On Deep Reinforcement Learning Algorithm

Cyclone is a common weather system.The weather with cyclones changes dramatically,often causing disasters such as heavy precipitation,thunderstorms,tornadoes,hail and so on.The monitoring of cyclone systems is one of the main concerns in meteorological disaster prevention and mitigation.Although the research on cyclones has achieved a lot of results,the variety of expressions and scales make the task of high-precision scientific analysis challenging,so it is necessary to improve the quality of existing algorithms.This paper focuses on large-scale cyclones in typhoon system and focuses smalland medium-scale cyclones in severe convection system.Methods based on deep reinforcement learning are proposed to locate the typhoon center on satellite images and interpret the windfield of severe convection on the Doppler radar radial velocity map.The main work is as follows:1.On the issue of typhoon center location,Weather satellites are major tools for monitoring round-the-clock typhoons.The research data consists of the image products of the geostationary meteorological satellite and the China Meteorological Administration tropical cyclone database.The proposed method turns the problem of typhoon center locating into an issue of searching for the typhoon center,and views the process of searching as a series of Markov Decision Processes.A deep reinforcement learning network consisting of convolutional neural network and deep Q network is trained to guide the decision in the search process.Meanwhile,typhoon discriminant rules are designed based on the output of the network to determine if a typhoon region is found during the process of searching for a typhoon center,so that the purpose of multiple typhoon detection and center location on a satellite image are realized.Experimental results demonstrate that the proposed method can complete the location process in about 12 steps and can effectively locate centers of typhoons in different forms with an average longitude error of 0.28 and a latitude error of 0.25.The precision of the proposed algorithm increases with the typhoon level and reaches100% at grades 5 and 6,and the average recall ratio reaches 91.6%.2.On the issue of windfield retrieval,Doppler weather radar is one of the main tools for monitoring severe convection weather in the inversion of strong convective wind field.Based on the understanding of the principle of radial velocity monitoring by Doppler radar,combined with the analysis of the mathematical model of the typical windfield mode of severe convection,a severe convection radial velocity map simulation system is established,which realizes the visualization of the typical convective windfield mapping onto the Doppler radial velocity map.A method which similar to "image restoration" for windfield retrieval of severe convection is proposed.The problem of windfield retrieval is decomposed into two parts: windfield initialization and windfield adjustment.In which the reinforcement learning is applied for preliminary tentative exploration.Experiments on the sample data generated by the simulation platform show that the proposed method can effectively interpret the simple composite windfield with the main field of cyclone,and prove that the similarity of the radial velocity map can represent the similarity of the vector flow field in the windfield adjustment phase.